SSL-LODDA: Self-supervised learning for low-light object detection with domain adaptation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Low-light object detection faces challenges from illumination degradation and noise.
• Limited annotated low-light data hinders supervised learning effectiveness.
• Significant distribution discrepancies exist between well-lit source and low-light target domains.

Purpose of the Study:

• To present a unified self-supervised domain-adaptive object detection framework.
• To address challenges in low-light object detection and cross-domain adaptation.
• To improve the robustness and scalability of object detection in adverse lighting.

Main Methods:

• Utilizes contrastive self-supervised learning (SSL) for illumination-invariant feature extraction.
• Employs Domain-Adversarial Neural Networks (DANN) and Maximum Mean Discrepancy (MMD) for domain alignment.
• Integrates Generative Adversarial Network (GAN)-based synthetic augmentation and pseudo-label refinement.
• Optimizes all components jointly using a multi-loss function.

Main Results:

• Achieves up to 82.3% mean Average Precision (mAP) on challenging low-light adaptation tasks.
• Consistently outperforms state-of-the-art domain-adaptive detectors across multiple datasets (PASCAL VOC, Cityscapes, MS COCO, ExDark).
• Ablation studies confirm the significant contributions of SSL, DANN, MMD, and pseudo-labeling.

Conclusions:

• The proposed framework offers an effective solution for scalable and robust low-light object detection.
• The joint optimization of self-supervised learning, domain alignment, and pseudo-labeling is crucial for performance.
• Demonstrates significant improvements in cross-domain adaptation for object detection under adverse lighting conditions.